quantum-projects

Imported from KRPFLWSTATE/quantum-projects (AGENTS.md). Install upstream with npx skills add KRPFLWSTATE/quantum-projects. Copyright stays with the author.

Learned User Preferences

• Never run git add, git commit, or git push until the user says the exact phrase "push all to github" (or an unambiguous equivalent).
• Never invent or fabricate experimental QPU results; report only observed jobs, counts, and metrics.
• After the decision-study completion freeze: "do a run" and "do a decision run" both execute exactly one decision-study SamplerV2 block (decision-study/study.py run-next --hardware) after frozen-protocol and live free-plan budget checks. The published campaign is COMPLETE at 6/6 jobs; further authorised runs remain blocked by the frozen job cap unless the author later freezes a new protocol in a separate directory. "do a legacy GHZ run" submits the old 3-qubit GHZ probe via dba-qpu-run/run_circuit.py then analyze_results.py. "resume the run" retrieves an existing decision-study attempt and never calls provider run again. Status, report, repair, documentation, and completion prompts are not hardware permission.
• Do not submit extra IBM jobs except when the user says to do a run, do a decision run, or do a legacy GHZ run.
• Do not modify the manuscript. Useful repository documentation may be added when the author explicitly requests a publication-repository pass.
• Do not overwrite Run 1 or Run 2 archives; do not commit save_credentials.py, .env, venv/, __pycache__/, or *_backup.json.

Learned Workspace Facts

• The IJCA/Taylor & Francis submission described in prior notes was rejected (letter dated 8 September 2026) as within scope but lacking sufficient contribution and clarity; do not treat that review-status text as current manuscript–repository alignment.
• Decision-study QAOA work lives under decision-study/ and must not be mixed into the 20-job GHZ archive under dba-qpu-run/results/runs/.
• Authoritative GHZ hardware archives live in dba-qpu-run/results/runs/; dba-qpu-run/qpu_run_telemetry.json is a last-run convenience copy.
• GHZ probe backend is chosen by service.least_busy(simulator=False, operational=True, min_num_qubits=3), not a hardcoded machine. Decision-study campaign backend is a frozen Open Plan pin (documented deterministic rule; not GHZ fidelity and not least_busy).
• Legacy GHZ SHA-256 baseline is captured once under decision-study/data/preservation/legacy_sha256.json; later commands must verify it, not silently rewrite it.
• Decision-study campaign archive status is COMPLETE (six physical jobs). Engineering study.py status still reports a submission blocker JOB_CAP_REACHED. Do not weaken that gate. Offline reproduction is python tools/reproduce.py and must not call Sampler run.
• Probe circuit is 3-qubit GHZ-style (H(0), CX(0,1), CX(1,2), measure_all()), 1000 shots, Hellinger fidelity over the union of all bitstrings without renormalizing leakage away.
• Python env is repo-root venv/ with tested pins in requirements-repro.txt and historical mins in requirements.txt; IBM credentials are local and gitignored.
• Run IBM scripts with unbuffered Python (PYTHONUNBUFFERED=1 / python -u); QiskitRuntimeService() / least_busy() can hang with empty buffered stdout.
• Frozen protocol hash: 0e8c4283c971f672853d3dfdf9167911eb834a0b3dc0aad16e1ef377f0cd046a. Do not format or refactor frozen decision-study/src or study.py in a way that changes those hashes.
• Post-collection analysis lives under tools/ with version publication-analysis-v2. Development installers live under archive/development/. Isolated offline reproduction is python tools/reproduce.py --output build/reproduction and must not copy into results/publication unless --refresh-publication is passed.